基于CNN-BiLSTM的中文微博立场分析研究

Research on Stance Detection in Chinise Micro-blog Based on CNN-BiLSTM

随着社交网络的迅速发展,产生了大量的社交网络文本数据。国际语义评测比赛SemEval-2016和自然语言处理与中文计算国际会议NLPCC2016均提出了针对社交网络文本进行立场检测分析的任务。传统的立场检测任务中,研究人员主要通过构建特征工程、情感词典等来挖掘文本语义特征,但需要花费大量人力在特征选择及其设计上。在深度学习中,长短时记忆网络LSTM可以获取句子的长时记忆信息,而一维卷积神经网络CNN能够获取文本的局部主要语义信息。文中提出一种基于词向量技术和CNN-BiLSTM的深度融合模型,首先利用卷积神经网络提取文本向量的局部特征,再运用双向LSTM网络提取文本的全局特征,解决了单卷积神经网络无法获取全局语义信息和传统循环神经网络梯度消失的问题。在NLPCC2016 Task4数据集上进行试验,实验结果有效提升了文本立场分类的准确率,验证了模型的有效性。

With the rapid development of social network,a large number of text data of social network have been generated. The tasks of position detection and analysis for social network text are proposed in SemEval-2016 and NLPCC2016. In traditional stance detection tasks,researchers mainly mine text semantic features by constructing feature engineering and affective dictionary,but they need to spend a lot of manpower on feature selection and design. In deep learning,long short-term memory(LSTM) can acquire long-term memory information of sentences,while one-dimensional convolutional neural network(CNN) can acquire local main semantic information of text. We propose a depth integration model based on word embedding technology and CNN-BiLSTM. Firstly,the local features of text vector are extracted by convolution neural network,and then the global features of text are extracted by BiLSTM network. The problem that single convolution neural network can not obtain global semantic information and the gradient disappearance of traditional cyclic neural network is solved. Experiments on NLPCC2016 Task4 data set show that the proposed model can effectively improve the accuracy of text position classification and validate the validity of the model.